The image needs to be cut apart and re-mosaiced after adjusting the colors in each piece. This can be done.
As an example, I extracted the green band of the image. To make my work simple (the computing platform I am using, Mathematica 9, does not easily extract pixels along arbitrary polylines), I rotated it to make some of the image boundaries perfectly horizontal and then focused on one of them in the middle of the image. I extracted the rows immediately above and below the apparent seam, shown in orange at the right of the figure below. (It comprises only 232 pixels, which is relatively limited information.) The "Profiles" plots in the middle of the figure show the intensities moving across this seam, with the profile lower in the image drawn in blue and the profile higher in the image drawn in red. The clear systematic difference in the two profiles reflects the relative lightness of the lower patch.
The scatterplot compares lower pixels to their nearest upper counterparts on log-log axes. This is the key idea: although there is substantial amount of scatter around a line, nevertheless a linear relationship is still apparent. By fitting a line appropriately to this scatterplot, we can recover a good estimate of the amount of correction to make in order to match the two sides.
Although many people would reflexively resort to using least squares to fit this line, that would be incorrect. The reason is that neither of the two patches should be favored as the "independent variable", but rather the two should be treated in a symmetrical fashion. Instead, the slope of the line, "gamma," was estimated in a robust way as the median ratio of the logarithms of the pixel intensities. The lower patch was gamma-corrected using this value (of 1.32), which darkened it, and re-mosaiced to the upper patch. The new "corrected" mosaic is shown at the left next to the original.
Even though the portions of the images used to compute the gamma correction did not overlap--they were separated by two pixels in the original mosaic--enough information remained in this cross-boundary comparison to do an adequate job of matching one patch to another. By choosing one reference patch and performing a sequence of gamma corrections across the full image, one could achieve noticeable improvement throughout.
This gamma correction can be applied separately to each band, finally reassembling the bands into a full-spectrum image.
Automating this process would require identifying the separate patches. I suppose these could be found with a clustering algorithm (supervised or unsupervised), but I have not implemented or tested one.
Although this approach is not perfect--starting with the original images and re-mosaicing them would be better--it shows potential to create a noticeable improvement, which at the least could be used as the starting point for any further efforts (such as the feathering mentioned in another answer).